Enhanced MRI brain tumor detection through fusion of transfer learning models and multi-classifier framework

Abstract Classification of brain tumors based on MRI images is an important task which introduces the benefit of early detection for brain tumor diagnosis, treatment planning, and it can positively affect patient's diagnosis. Early detection of brain tumors can improve survival rates and quality of life when diagnosis occurs in the early stages and is followed by timely treatment planning. While MRI images can have a good diagnostic capacity for brain tumors they need a radiologist to interpret and classify them which will have a variation and delay for treatment planning, this paper presents a fully automated framework using deep learning to classify multi-class brain tumor from MRI images with less error. A deep learning model based on a transfer learning mechanism and early fusion technique is proposed in this work that fuses heterogeneous nature models. Deep feature representations are extracted from four different Transfer Learning architectures namely: InceptionV3, Xception, ResNet50, and VGG16 and fused at the early stage using concatenation followed by a fully connected meta-classifier to present the final classification into four types of tumors: glioma, meningioma, pituitary tumor, and no tumor. To support a reliable evaluation of the designed method the experiments are done using K-fold cross validation. The experiments showed that the proposed fusion based model performs better than standalone models and reached 99.42% accuracy, 99.28% precision, 99.42% recall, and a very low loss of 0.015 on one of its configuration where the standalone networks such as ResNet50 performance level is much lower. These results confirm that integrating complementary representations from multiple deep learning models enhances classification accuracy and generalization. On the datasets evaluated here, the framework offers an efficient and promising basis for automated screening; broader multi-institutional and patient-level validation is still required before it could be expected to reduce dependency on manual interpretation or to support clinical decision-making.